FBPConvNet - tensorflow

http://ieeexplore.ieee.org/document/7949028/

This is tensorflow implementation for Deep Convolutional Neural Network for Inverse Problems in Imaging, TIP (2017).

• applications (forward model with shift-invariant normal operator):
• 2D sparse-view CT reconstruction
• reconstruction of accelerated MRI
• Deconvolution of shift-invariant

Whole codes are forked and modified from https://github.com/jakeret/tf_unet.

Training configuration

• Tensorflow 1.1.0
• 1 or 2 GPUs (TITAN X pascal arch.)
• MacOS X 10.12.6
• Python 2.7.12

Data - XYCN format (ellipsoids, downsampling factor : x 20)

• train : https://drive.google.com/open?id=1FTOgM2vOQaGSokEDtOaPNdBTto6h5yFi
• test : https://drive.google.com/open?id=1w_kPao6L2UwhTKIgcr_3o62A6vYYtX_r
• If you want to make fbp images, you can find file_generator in tf_unet/layers.py (load_whole_data function.)

illustration

Commands

Before starting,

pip install pillow matplotlib scipy scikit-image h5py

To start training a model for FBPConvNet:

python main.py --lr=1e-4 --output_path='logs/' --train_path='train_data/*.mat' --test_path='test_data/*.mat' --features_root=32 --layers=5 

To deploy trained model:

python main.py --lr=1e-4 --output_path='logs/' --train_path='train_data/*.mat' --test_path='test_data/*.mat' --features_root=32 --layers=5 --is_training=False

You may find more details in main.py.

Contact

kyonghwan.jin@gmail.com